Research On Feature Fusion And Behavior Recognition Of Human Activity

Human activity recognition is for the person in the video motion analysis, through the automatic tracking of human behavior information of the whole or local information, to identify the classification and understanding of human behavior. In intelligent video surveillance, human-computer interaction, content-based video retrieval and robotics and other fields have a wide application prospect. To the human behavior recognition of interest points feature extraction, multi feature fusion, multi-scale fusion and multiple kernel learning classification recognition method, such as technology research, puts forward the improvement and innovation, obtains a better recognition effect, the main innovation achievements obtained in this paper are as follows:1. A human activity recognition method based on the fusion of adjacent spatial interest points is proposed. The method for interest points feature that contains only local information and adjacent to the lack of relevant information, in the feature fusion of design, the external appearance of the human body static characteristics, sports interest point dynamic characteristics and the motion velocity of the three kinds of feature fusion, by KTH human action data set the ratio. The experimental results show that the multi feature fusion of improved method of velocity behavior with high sensitivity and enhanced the distinction between the samples, and improve the recognition precision of human behavior.2. A human activity recognition method based on scale mixing feature is proposed. Appearance outline of the method for human motion cycle characteristics and behavior of interest points distribution characteristics, the design of the interest points feature is extracted in different scales, combining the characteristics of different scales and human appearance characteristics, by KTH human action data set experiment. Experimental results show that the scale features is introduced in this paper can improve the discrimination of different motion behavior, reached the point to improve the accuracy of the recognition of human behavior.3. A study on human activity recognition method based on multi kernel learning MKL is proposed. The method for in multi-scale and mixed use in each kind of scale in different behavior in different contribution rate, single kernel SVM classification method is difficult to distinguish differences in different scales, the different scale design their own weight coefficient proposed multiple kernel learning method for classification and recognition method based on, and in the KTH human action data set were compared. The experimental results show that the introduction of multiple kernel learning model and optimization method of kernel weight in human action recognition plays a very good efffect, to improve the accuracy of the recognition of human behavior.